GuidesRunning & maintaining

OpenClaw "runtime owner was not published": agent reads, never replies

OpenClaw agent silent while health stays green? The log line, what triggers it, the restart that clears it, and the 2026.11.5 fix.

November 17, 2026The Everpod team
The short answer

“prepared reply dispatch runtime owner was not published for main” (with your agent’s id in place of main) means OpenClaw failed to finish preparing that agent’s model setup, usually after a preparation step ran past its time limit, and then refused every reply for the agent instead of trying again. Messages arrive, the typing indicator may keep going, nothing is sent back, and the Gateway’s health check stays green. On releases before 2026.11.5 only a Gateway restart clears it: openclaw gateway restart, then resend what went unanswered. OpenClaw 2026.11.5 carries the fix, which rebuilds the agent’s setup when the next message arrives.

The lines that identify it

The refusal is logged once per message the agent fails to answer. Earlier in the log, in the reports, there is a line saying a preparation step timed out or was superseded. The forms quoted in OpenClaw’s issue tracker since September 2026 include:

prepared model runtime publication (workspace plugins; agent main) timed out
auth-triggered model runtime refresh failed: Error: prepared model runtime publication (...) timed out
prepared model runtime publication was superseded for <agent folder>
prepared reply dispatch runtime owner was not published for main
Copied

In 2026.11.4 and 2026.11.5 the last line can go on: “… was not published for main; recovery failed. Check Gateway logs for the underlying publication error.”, or “recovery is waiting for retry” in place of “recovery failed”. Only one agent may be affected: on a Gateway with several, the others go on answering.

To look for it, read the Gateway’s log through the CLI, which works as long as the Gateway process is up:

openclaw logs --limit 2000 --plain | grep -E 'runtime owner was not published|runtime publication'
openclaw logs --follow
Copied

The second command shows new lines as they arrive; send the agent a message while it runs and the refusal appears for it. The logs reference gives the file itself as openclaw-YYYY-MM-DD.log in the logs/ folder of the state directory. In Docker, docker compose logs -f openclaw-gateway from the Compose directory shows the same lines.

Why the health check stays green

OpenClaw’s health reference says /healthz confirms only that “The HTTP server is live”, and the stricter /readyz checks startup, draining, agent databases and channel accounts, not whether each agent can reply. Reporters with this failure found every probe passing: the HTTP endpoint, RPC, the Telegram channel’s own check. A Docker healthcheck or an uptime monitor pointed at /healthz will not restart the container, so the agent can stay silent for a day or more; one report counted 32 hours.

That is also what separates it from an agent silenced by its model account: there, every turn fails with an error from the model provider, and switching models in chat brings it back. Here no request reaches the provider at all, so /model changes nothing.

What sets it off

The shared ingredient in the reports is a moment when preparing an agent’s models takes longer than OpenClaw allows, which maintainers and reporters put at 120 seconds. Reported triggers, on macOS, Linux and Windows, from 2026.9.5 to 2026.10.6:

A contributor reading the code before the fix found why it never heals: the time limit stops waiting for the preparation without stopping it, and the failure is kept, so every later message for that agent meets the same refusal.

Getting the agent answering again

openclaw gateway restart
# in Docker, from the Compose directory:
docker compose restart openclaw-gateway
Copied

A restart rebuilds every agent’s setup from scratch, and reporters found the agent answering normally after it. The messages it ignored are not answered by themselves: resend them. If other agents on the same Gateway are mid-task, openclaw gateway restart --safe waits for their work to finish first, for up to five minutes, as the restart reference describes; the restart guide covers the service managers behind it. If the timeout line comes back during the restart’s own startup, that is the slow-start case above.

The fix in 2026.11.5

The fix (pull request #181300) was merged in October and first reached a stable release in 2026.11.5, published on November 17, 2026; 2026.11.4 and earlier do not have it. With it, the next message for a failed agent tries to rebuild its setup instead of being refused, and the typing indicator stops while it waits. When the maintainers closed #158592 that day, they asked anyone still seeing the message on 2026.11.5 or later to reopen it with the log lines.

2026.11.5 had its own open reports in its first days, including updates stopping partway, so read its entry in the latest-version guide before taking it, make a backup, and run the update from a terminal as the update guide describes. Until you update, the restart above is the recovery each time it happens.

On 2026.11.5, if two rebuilds in a row fail, the message is set aside in the failed queue with the reason runtime-recovery-failed, keeping its text. OpenClaw’s troubleshooting page says to look in openclaw logs --follow for the publication error underneath, correct the model or plugin problem it names, and resend the message.

Hearing about it on an unattended server

Since the health check cannot see it, a server you do not watch needs a check of its own. A script run every few minutes from cron that looks for the line and mails you:

#!/bin/sh
# Alert when an OpenClaw agent has stopped replying with "runtime owner was not published".
hits=$(openclaw logs --limit 500 --plain | grep 'runtime owner was not published')
if [ -n "$hits" ]; then
  printf '%s\n' "$hits" | tail -n 5 | mail -s "openclaw: an agent is not replying" you@example.com
fi
Copied

Replace the address and the mail command with whatever reaches you. The check looks at the last 500 log lines, so after a restart it keeps alerting until newer lines have pushed the old ones out of that window. It sees only this failure; an agent that is silent for another reason needs the triage in the not-responding guide.

Give your AI agents a permanent cloud computer.

We set it up and secure it, so all you do is use them. Run Claude Code, Codex or any agent you use, more at once than your laptop can hold, and keep them working when it’s closed. Or have OpenClaw set up and run for you.

See the developer pod

From $24/mo · your own plan or API key

Get your OpenClaw agent

$29/mo, cancel anytime · model usage included